Machine learning of projected 3D shape

Abstract

This thesis primarily investigates the potential of the Pairwise Geometric Histogram (PGH) representation as the basis of a machine learning edge and view-based 3D object recognition computer vision system. The work extends 20 years’ worth of associated research within the TINA computer vision research group [1]. PGHs have formerly been engineered as a solution to the presented problem, directly addressing all of the invariance characteristics required by such a representation. Previous research has proven the power of the proposed techniques for 2D object recognition through difficult, real-world viewing conditions including scene clutter and occlusion. This project extends the associated methodologies into the third dimension, exploring methods for representing scaled 3D objects’ continuous appearances around their view-spheres. The research agenda has also included a comparative analysis of the pre-existing TINA [1] stereo vision-based 3D Model Matching (3DMM) system, which is able to localise specified 3D objects in 3D scenes. In support of both mono and stereo methodologies, a quantitative scheme for accurately localising and verifying the presence of hypothesised image-projected 3D edge-feature models has been implemented. Full view-sphere sampled 3D model matching tests have been conducted for the competing methodologies, identifying significant shortfalls with the stereo-based approach to 3D model matching. The more powerful and reliable view-based techniques are subsequently analysed with regard to the more demanding task of comparative 3D object recognition. Institution The University of Manchester Candidate Simon Coupe Degree Title Doctor of Philosophy Thesis Title Machine Learning of Projected 3D Shape Date 30th September 2009